Rabbit's R1 can now learn tasks by watching you

When the bright orange Rabbit R1 arrived earlier this year, reviewers pulled no punches. Engadget famously branded it "a toy that fails at almost everything," with its headline Large Action Model (LAM) mostly missing in action.

After rolling out over 20 software patches since the spring, Rabbit is finally pushing its most critical update yet. Starting today, every R1 owner gets beta access to teach mode, a feature designed to let users train the AI to handle web tasks automatically.

Recording clicks to automate the web

Teach mode essentially works like an interactive macro recorder hosted on desktop browsers through Rabbit's Rabbithole portal.

You log in, open the target website, and input your credentials if needed. CEO Jesse Lyu notes that Rabbit avoids storing raw usernames and passwords, choosing instead to save session cookies for the R1 to execute tasks later. Security remains top of mind here, especially since Rabbit had to patch a major data exposure vulnerability back in June.

Once you name and describe your automation, you simply perform the workflow. The software converts each click and input into executable steps. During a demo with Engadget, Lyu taught his R1 to post a tweet.

While Rabbit claims no coding knowledge is required, an engineering mindset definitely helps. Users can inspect the captured workflow, annotate individual steps, and replay the automation on their computer before sending it to the pocket device.

Running it on the actual R1 is still a bit clunky. The device talks aloud through every single step, and execution takes noticeable time. Rabbit claims this vocal playback is an intentional design choice requested by early testers to track progress.

Multi-platform automations and an agent store

Lyu openly concedes that using an R1 for simple, single-destination tasks is often slower than tapping an existing smartphone app.

Instead, he sees teach mode shining across messy, fragmented workflows. One example he offered involves snapping a photo of a handwritten grocery list with the R1's camera, then having the device automatically order those items from local stores online. Another pitch is bridging competing ecosystems, like wrangling Apple HomeKit and Google Home without waiting for standard protocols.

Rabbit also envisions an eventual economic layer. Community automations are currently free on Rabbithole, but Lyu plans to introduce a marketplace where creators can sell automations while Rabbit takes a cut, mirroring mobile app stores.

Down the road, Lyu wants teach mode to navigate desktop software like Excel, claiming this path will eventually yield an artificial general intelligence that understands every human interface. When asked about anti-bot measures like CAPTCHAs, Rabbit argued verification systems will need to evolve to recognize authorized, user-driven agents.

Why it matters

For existing R1 owners who left the gadget in a drawer, teach mode is the first real sign of the device Rabbit originally promised. But if you do not already own one, we think this update is far from a reason to buy.

Training web automations requires patience and troubleshooting. For everyday users, opening an app on an iPhone or Android device will still be ten times faster than recording, testing, and debugging a cloudy web script on Rabbithole.

For the broader industry, Rabbit is highlighting a genuine frontier: navigating interfaces on behalf of users when closed APIs refuse to talk to each other. Taking an agent-first approach to web actions sidesteps developer roadmaps entirely.

However, the underlying mechanics carry massive risks. Relying on saved browser cookies creates a tempting target for bad actors, particularly for a startup with past security stumbles. Furthermore, brittle web frontends change constantly, meaning automations are bound to break without warning.

In our view, rebranding basic robotic process automation (RPA) as the path to AGI feels like a stretch. Rabbit has delivered a fun playground for enthusiasts, but until major platform holders build native agentic controls directly into operating systems, teach mode remains a fragile workaround.